Abstract
This paper explores a novel clustering approach for multimodal Glioblastomas (GBM) characterization using the magnetic resonance image (MRI) modality. We define our segmentation problem as a linear mixture model (LMM). In every segmentation process, we generate a non-negative matrix with GLCM features from every MRI slice and a rank-two NMF (Non Negative Matrix Factorization) is applied.
Our method process in four levels of segmentation. In the first one, the LMM matrix for the whole brain was generated from FLAIR modality to extract whole tumor region, which considered as the region of Interest (ROI). In the second level, we extract the ROI from T1c modality and the LMM matrix was generated from only this ROI to extract necrosis region. The principle will be the same for the other two levels to extract the enhanced and the non-enhanced region. Quantitative and qualitative assessment over the publicly dataset from MICCAI 2015 challenge (BRATS 2015) demonstrated that the proposed method could generate a competitive efficiency for high grade Glioblastomas characterization among several competing method.
In order to highlight the performance of our method, we propose a comparative study with unsupervised segmentation methodologies (K-means, fuzzy C-means (FCM), gaussian mixture model (GMM) and hierarchical non-negative factorization (hNMF)) over the publicly BRATS 2015 dataset by computing validation metrics (the sensitivity, the dice and the specificity). The obtained results could attest the performance of the proposed algorithm compared to the unsupervised segmentation methodologies.